Development and internal validation of a multimodal deep learning model for screening-level lower-limb alignment using plantar 3D foot scans

Lower-limb alignment is a key determinant of musculoskeletal loading and is associated with biomechanical risk pathways for conditions such as knee osteoarthritis and plantar fasciitis. However, routine alignment assessment still relies on radiography or specialist clinical examination, limiting scalability for early screening. Here, we investigated whether plantar foot morphology captured by weight-bearing 3D scanning can serve as a computational screening feature for proximal lower-limb alignment, and we developed a multimodal deep learning framework to support scalable alignment screening. We established a cross-sectional dataset of 167 participants (adults and children aged 6–60 years) with standardised plantar 3D foot scans and clinical labels for arch type, hindfoot alignment, and frontal-plane knee alignment. The proposed model jointly learns from plantar scan images and 16 geometric morphological features, integrated using feature-wise attention to prioritise alignment-relevant cues across modalities. Performance was evaluated using accuracy, precision, recall, F1 score, and AUC on a held-out test set, and interpretability was assessed with class activation mapping (CAM). Ablation studies compared alternative image encoders and fusion variants, including image-only, morphology-only, and fusion without attention. The attention-based multimodal model showed competitive but task-dependent performance, achieving accuracies of 89% for arch type, 81% for hindfoot alignment, and 76% for knee alignment. The image-only ViT achieved the highest accuracy for arch classification, while the ViT without-attention configuration achieved slightly higher accuracy and macro-F1 for knee alignment. These findings indicate that the benefit of multimodal attention fusion varied across alignment tasks. Activation maps consistently highlighted clinically relevant plantar regions associated with alignment patterns, suggesting morphological correlations consistent with a distal-proximal kinetic chain hypothesis. Together, these findings demonstrate the potential of plantar 3D scanning combined with multimodal, explainable deep learning as a non-invasive approach for screening alignment-related biomechanical risk factors, warranting further external validation before clinical translation.